Am. J. Trop. Med. Hyg.,95(1), 2016, pp. 63–69 doi:10.4269/ajtmh.15-0647
Copyright © 2016 by The American Society of Tropical Medicine and Hygiene
Landscape Fragmentation as a Risk Factor for Buruli Ulcer Disease in Ghana
Jianyong Wu1,2* and Erica A. H. Smithwick1
1Department of Geography, The Pennsylvania State University, University Park, Pennsylvania;2Gillings School of Global Public Health,
The University of North Carolina, Chapel Hill, North Carolina
Abstract. Land cover and its change have been linked to Buruli ulcer (BU), a rapidly emerging tropical disease. However, it is unknown whether landscape structure affects the disease prevalence. To examine the association between landscape pattern and BU presence, we obtained land cover information for 20 villages in southwestern Ghana from high resolution satellite images, and analyzed the landscape pattern surrounding each village. Eight land-scape metrics indicated that landland-scape patterns between BU case and reference villages were different (P< 0.05) at the broad spatial extent examined (4 km). The logistic regression models showed that landscape fragmentation and diver-sity indices were positively associated with BU presence in a village. Specifically, for each increase in patch dendiver-sity and edge density by 100 units, the likelihood of BU presence in a village increased 2.51 (95% confidence interval [CI] = 1.36–4.61) and 4.18 (95% CI = 1.63–10.76) times, respectively. The results suggest that increased landscape fragmenta-tion may pose a risk to the emergence of BU.
INTRODUCTION
Landscape disturbances alter ecosystem patterns and pro-cesses, and are increasingly recognized to have cascading effects on ecosystem functions at local to global scales.1 Spe-cifically, human-driven land cover changes that cause habitat loss and fragmentation have been associated with the out-break and transmission of multiple infectious diseases.2,3For example, in the Amazon rainforest, deforestation was associ-ated with increased malaria prevalence due to an increase in suitable habitat of the malaria vector, Anopheles darlingi.4 The construction of dams in Senegal was found to be respon-sible for the rise of the outbreaks of schistosomiasis due to increases in water habitat beneficial to the vector (snails) and parasite transmission.5In the northeastern United States, Lyme disease risk is associated with forest fragmentation, largely through resultant modification of trophic interactions that favor vector transmission.6 These studies highlight an increased understanding of landscape patterns and disease risk, but to date few studies have evaluated the usefulness of pattern metrics on Buruli ulcer (BU) prevalence,7despite the fact the changes in land use have been previously reported in association with the disease.8–10Understanding how landscape patterns influence human disease could aid the development of landscape-level management plans to reduce disease risk.
BU is a skin infection caused byMycobacterium ulcerans, and has emerged in over 30 countries worldwide and become the third most common disease caused by Mycobacteria, after tuberculosis and leprosy.11 Though its transmission mode is still unclear, it is generally recognized to be associated with aquatic habitats11 and aquatic biting insects12 as well as nonbiting aquatic invertebrates11although results are equiv-ocal.13 Other landscape disturbances, such as deforestation and agriculture, have also been linked to BU prevalence. In Benin, the dynamics of BU disease were correlated with human alterations to landscapes and natural land cover.9 Recently, in Ghana, Wu and others10 documented positive relationships between BU prevalence and differences in
min-ing and agriculture between endemic and nonendemic regions. However, it is still unclear whether these changes in habitat loss were indicative of changes in fragmentation and land-scape structure, which may have differential effects on eco-system processes14that influence disease emergence.
Ghana is one of the most prevalent countries of BU dis-ease, second only to Cote d’Ivoire.15 Rapid land cover change has taken place in Ghana in recent decades. For example, forest area has decreased roughly 2% per year since the 1990s,16while agriculture has grown very rapidly and expanded at an annual rate of 5.5%.17 In addition, across Ghana, gold mining has increased sharply in recent decades. From 2004 to 2009, overall gold production in Ghana increased from 2.6% to 3.8% of global production,18 which poses risks to environmental and human health.19
In this study, we explore whether landscape patterns can be used as an indicator of BU disease, given its importance as a rapidly emerging tropical disease. Specifically, two ques-tions are answered: 1) do landscape metrics for landscape fragmentation and diversity differ between BU case and reference villages and 2) to what degree is BU presence at a village level associated with landscape metrics surrounding villages? We hypothesized that BU case villages would have higher levels of fragmentation and that the correlation would be positive at broader spatial extents (e.g., 4 km away from village center) if landscape heterogeneity increased.
METHODS
Study area.Our study area included 20 villages in south-western Ghana, which were located in three study regions (Figure 1). We selected two study regions (hereafter called Subin and Ayanfuri) in a BU endemic area where BU cases were clustered.10 A reference region (Kedadwen) was also included, where BU has not been reported. These three regions were selected after the discussion with a diverse team of experts during an interdisciplinary workshop in 2008 and then were confirmed by a field visit. The key characteristics of these study areas, such as climate, the type of vegetation and land cover, and geology, are similar or comparable (Supple-mental Table 1). The area of each study area is near 580 km2. Six villages from Subin region (V1–V6) and five villages from Ayanfuri region (V7–V11) were selected as the BU case *Address correspondence to Jianyong Wu, Gillings School of Global
Public Health, The University of North Carolina, Chapel Hill, NC 27599. E-mail: [email protected]
villages. Five villages from Kedadwen region (V16–V20) were selected as the reference villages. In addition, two villages in which BU cases have not been reported were selected from the Subin (V12 and V13) and Ayanfuri regions (V14 and V15), respectively, as reference villages. These villages were roughly randomly selected after the distance between villages and the number of villages in these areas was taken into account. In total, 11 villages with BU cases and nine villages without BU cases were used.
BU disease data.Hospital-based BU data were collected between 2007 and 2010. When patients with skin infections visited hospitals and clinics, they were examined by experi-enced doctors based on the symptoms of the infection. These cases were confirmed by clinical diagnosis rather than labo-ratory testing of pathogens because of technical difficulties. Once a case was confirmed, the age, gender, residence, and other information were recorded. Therefore, the village in which a patient is living was identified.
Land cover and landscape pattern analysis.RapidEye sat-ellite imagery with a spatial resolution of 5 m acquired in January 2012 was used to obtain land over information in the study areas. The images were classified into six types of land cover classes (urban, mining area, water, grassland,
of calculation of these metrics can be found elsewhere.21 FRAGSTATS v.4.2 (Amherst, MA) was used for all met-rics calculations.21
Statistical analysis.To test for the difference in individ-ual landscape metric between BU case and reference vil-lages (question 1), we used one-way analysis of variance (ANOVA). Multivariate analysis of variance (MANOVA) was conducted to test the overall difference in landscape patterns between two groups of villages. Before conducting ANOVA and MANOVA, we validated some assumptions for these analyses, including normality, homogeneity of vari-ances and covarivari-ances, linearity, and outliers of dependent variables. Because some variables were highly correlated (Pearsonr> 0.9,P< 0.01), we only put one of these highly correlated variables in the model for MANOVA. Therefore, three dependent variables, ED, ENN_SD, and SHDI were included in MANOVA. To reduce type I error, we adjusted multiple comparisons with a Bonferroni correction. The same methods were used to test the difference in land cover components between two groups of villages. The relation-ship between land cover classes was examined with Pearson correlation analysis. To examine BU presence in villages as a function of landscape metrics, buffer radius, and land cover (question 2), we used logistic regression models. The dependent variable was BU presence in a village, which followed a binomial distribution. Specifically, if a village had BU cases, the value was set to one, otherwise, the value was set to zero. First, univariate logistic regression models were used to examine the association between BU presence and an individual variable. Then multivariable logis-tic regression models were used to examine the association between BU risk and the multiple variables simultaneously. Before carrying out multivariable logistic regression, the col-linearity between independent variables (landscape metrics) was examined by Pearson correlation analysis and vari-ance inflation factors (VIF). If the collinearity was identified (e.g.,r> 0.6,P< 0.01, or VIF > 5), only one of these corre-lated variables was put into the model. The fit of model was evaluated using Akaike’s Information (AIC) as the cri-terion, which assumes the model is better fitted if the value of AIC is smaller.10The association between the likelihood of BU presence in a village and landscape metrics was assessed by an odds ratio and 95% confidence intervals, calculated from the logistic regression models. Since the ranges of the metrics are different in magnitude (e.g., PD ranged from 332 to 1,278, while SIDI ranged from 0.39 to 0.74), the effects of the increase by one unit in these metrics are not comparable. To address this issue, we normalized the metrics to obtain a reasonable scale,22 for example, SIDInew = SIDI × 10. In addition, to evaluate how sensitive the models to the reference village selection, univariable logistic regression models were used to examine the associ-ation between the presence of BU and landscape metrics in BU endemic areas that included nine BU case villages and four reference villages.
RESULTS
In total, 73 cases were confirmed in 11 villages between 2007 and 2010. Among these cases, 40 cases were male and 33 cases were female, 34% of total cases were young peo-ple (0–19), 44% were adults (20–60), and 22% were older
people (60 and above). The highest number of cases was reported in Dunkwa (33 cases), followed by Ayanfuri, Subin, and Nkonya (14, nine, and five cases, respectively). Pokukrom, Nkotumso, and Nyinawusu reported four cases in each village. The remaining villages only reported one or two cases (Supplemental Figure 2).
Land cover components between BU case villages and reference villages were significantly different (Figure 2 and Supplemental Table 2). At 4 km, except the percentage of agricultural area, the percentages of other five types of land cover classes were significantly different between BU cases villages and references villages (P < 0.01). The differences in land cover components between two groups of villages at 2 km were similar as those observed at 4 km, except the dif-ferences in the percentages of urban, mining, and forest areas were significant at the level of 0.05, instead of 0.01. However, at 1 km, the differences in land cover components between two groups of villages were only significant for the percentages of water area (P < 0.05) and grassland (P< 0.01). In the BU case villages, the percentage of agri-culture was the dominant land cover class at all buffer distances, followed by grassland, forest, urban, water, and mining. As the buffer distance increased, the percentages of agricultural area and forest increased, while the percent-ages of urban area and grassland decreased. In the reference villages, the predominant land cover classes differed signifi-cantly with buffer distance. Agriculture was the predominant land cover class at 1 and 2 km, while forest became the pre-dominant land cover class at 4 km. Pearson correlation anal-ysis of the percentages of land cover classes at 4 km indicated that strong positive correlations (P< 0.01) existed between water, grassland, and mining areas, while negative correla-tions existed between forest and urban, and between forest and agriculture (Supplemental Table 3).
Landscape patterns were significantly different between BU case villages and reference villages at the radii of 4 and 2 km, but not significant at the radius of 1 km. At 4 km, PD, ED, ENN_MN, and ENN_SD were different between BU case and reference villages (P < 0.05). SHDI, SHEI, SIDI, and SIEI also differed between BU case and reference villages (P< 0.01). At 2 km, only SHDI, SHEI, SIDI, and SIEI were significantly different, while no metrics were sig-nificantly different between BU case villages and reference villages at 1 km (Table 1 and Supplemental Table 4).
Univariate logistic regression showed PD, ED, and four diversity indices (SHDI, SHEI, SIDI, and SIEI) were posi-tively associated with BU across all villages (Supplemental Table 5). The similar associations were obtained when only villages in endemic areas were included in the models (Sup-plemental Table 6). In the multivariable logistic regression model, the above metrics had significantly positive associa-tions with the presence of BU after controlling buffer radius and the percentage of forest area. Specifically, model results indicated that an increase in PD and ED by 100 units were associated with the increases in likelihood of BU presence in a village by 2.51 and 4.18 times, respectively. When diversity indices increased by 0.1 units, the likelihood increased 2.50– 10.92 times (Table 2).
DISCUSSION
We compared land cover and landscape patterns in BU-endemic areas with nonBU-endemic areas and examined the 65
association between landscape pattern and BU using high resolution satellite images at multiple spatial extents. Our results indicate that increases in land cover fragmentation and landscape diversity were associated with the presence of BU disease in a village in Ghana. The finding is consis-tent with recent studies highlighting the close relationship between human health and landscape changes due to distur-bance.1,10,23 Although the relationship between land cover and BU disease has been examined in several studies,9–11
the comparison of fine-scale landscape patterns and associ-ated metrics between endemic and nonendemic areas has rarely been presented.7
It is increasingly recognized that landscape structure plays an important role in disease transmission.24–26 In terrestrial systems, environmental changes driven by human activities affect the form, size, and connectivity of patches in a land-scape, leading to habitat fragmentation, which is considered as a health risk because it may alter the density, abundance, FIGURE2. Comparison of land cover classes in villages with Buruli ulcer (BU) and without BU (the error represents 95% confidence inter-val). **P< 0.01, *P< 0.05.
TABLE1
Comparison of landscape pattern metrics in BU case villages and reference villages with one-way ANOVA at different spatial scales
Variable Mean ± SD (N= 60)
4 km (N= 20) 2 km (N= 20) 1 km (N= 20)
Fvalue P Fvalue P Fvalue P
PD 683 + 259 4.515 0.048 2.260 0.150 0.690 0.417
ED 588 + 139 6.112 0.024 2.810 0.111 0.640 0.436
ENN_MN 24.38 + 2.88 5.150 0.036 0.580 0.455 0.180 0.677
ENN_SD 30 + 8 6.540 0.020 0.610 0.443 0.070 0.797
SHDI 1.20 + 0.21 18.820 < 0.001 7.520 0.013 3.470 0.079
SHEI 0.62 + 0.11 19.460 < 0.001 7.430 0.014 3.000 0.100
SIDI 0.62 + 0.08 8.720 0.009 6.130 0.024 3.520 0.077
SIEI 0.73 + 0.10 8.850 0.008 6.370 0.021 3.400 0.082
and geographic distribution of hosts, vectors, and pathogens involved in disease transmission and change the ecology of microorganisms,27 as has been shown for Lyme disease in the northeastern United States.25As another example, habi-tat fragmenhabi-tation was associated with the increase in hanta-virus hosts, potentially increasing the potential for an outbreak of hantavirus infection in Panama.27 Our results reveal that several landscape pattern metrics, for example, PD, ED, SHDI, and SIDI, were significantly different between BU case villages and reference villages, and two landscape frag-mentation indices, PD and ED, were significantly and posi-tively associated with BU presence in a village.
There are two primary reasons that fragmentation may be associated with BU. First, landscape fragmentation increases the probability that vectors in the preferred habitat encoun-ter habitat edges, which results in a higher chance of the vectors moving into other habitats, facilitating disease transmission.26 Second, the increase in landscape fragmentation may enhance human exposure to the disease, as evidenced by the effect of forest fragmentation on Lyme disease.24Specifically, the high resolution satellite images revealed that many mining activi-ties existed in the endemic areas, contributing to landscape fragmentation24; given the large number of people working and living near mining communities or small scale farms, the landscape structure of these environments may contribute to increased exposure to the bacterium that causes the disease. Both of these ideas will require further research.
Scale is an unavoidable issue in ecology because it influ-ences the results of landscape pattern analysis.28 In this study, we analyzed landscape pattern around each village at three spatial extents (radii at 1, 2, and 4 km), which were also used in other studies to analyze landscape context.26 In a study on the relationship between BU and landscape met-rics in Benin, radii from 0.8 to 2 km were used,7which were also included as radii in this study. Our results indicated that eight metrics were significantly different between BU case villages and reference villages at the distance of 4 km, four metrics were significantly different at the distance of 2 km,
while none were significantly different at the distance of 1 km, suggesting the importance of landscape heterogeneity and context at broader spatial extents, consistent with the result from a previous study.29 Collectively, the MANOVA suggested results were different at 4 and 2 km but not signifi-cant at 1 km. Both mean and standard deviation of land-scape metrics declined with increasing radii, indicating that the difference was not caused by changes in variance with scale (Supplemental Table 3). Moreover, the landscape met-rics are calculated at patch, class, or landscape levels, not for individual pixels, constraining the influence of increased sampling units with scale. Finally, the multivariate regression indicated a greater influence of landscape metrics and land cover than distance, per se.
Ultimately, understanding the effects of landscape struc-ture on disease transmission can help to predict and control disease outbreaks. To date, the use of land cover and land-scape pattern metrics as predictors for BU presence or risks has been explored.7,9,30However, studies on the relationship between landscape diversity and BU disease are still rare. Brou and others8 reported that landscape diversity was related to BU in Cote d’Ivoire. They found that the BU risk zones were located at irrigated rice field cultures areas, banana fields, and areas close to dams used for irrigation. However, their study did not measure any landscape diver-sity index and quantify the relationship between BU and the landscape diversity index. Campbell and others7 examined three landscape metrics with BU risks but did not find posi-tive associations, which could be explained by several rea-sons. First, Campbell and others examined the landscape at relatively small scales (radii ranged from 0.8 to 2 km). Our study showed that stronger associations were observed at 4 km. Second, Campbell and others used Landsat ETM + imagery with the spatial resolution of 30 m, which is rela-tively coarse and not favorable for landscape pattern analy-sis. In addition, Campbell and others only chose a limited set of metrics (three). Our study used higher resolution imagery, broader spatial extents, and a broader set of metrics, and TABLE2
Multivariable logistic regression analysis of the association between BU risk and landscape metrics in southwestern Ghana
Model ID Variables β
OR
P AIC
Point estimate 95% CI
1 PDnew 0.92 2.51 1.36 4.61 0.003 54.32
Forest −0.12 0.89 0.83 0.95 < 0.001
Buffer radius 0.63 1.88 0.95 3.72 0.070
2 EDnew 1.43 4.18 1.63 10.76 0.003 56.62
Forest −0.10 0.90 0.86 0.95 < 0.001
Buffer radius 0.59 1.80 0.93 3.49 0.083
3 SHDInew 0.91 2.50 1.52 4.09 < 0.001 52.12
Forest −0.08 0.93 0.89 0.97 0.001
Buffer radius 0.35 1.42 0.77 2.63 0.264
4 SHEInew 1.81 6.13 2.28 16.49 < 0.001 52.27
Forest −0.08 0.93 0.88 0.97 0.001
Buffer radius 0.37 1.45 0.78 2.69 0.240
5 SIDInew 2.39 10.92 2.84 41.98 0.001 52.30
Forest −0.09 0.91 0.87 0.96 < 0.001
Buffer radius 0.38 1.46 0.79 2.68 0.226
6 SIEInew 2.09 8.10 2.48 26.44 0.001 52.07
Forest −0.09 0.91 0.87 0.96 < 0.001
Buffer radius 0.39 1.48 0.80 2.74 0.207
AIC = Akaike’s information criterion; BU = Buruli ulcer; CI = confidence interval; ED = edge density; OR = odds ratio; PD = patch density; SHDI = Shannon’s diversity index; SHEI = Shannon’s evenness index; SIDI = Simpson’s diversity index; SIEI = Simpson’s evenness index. PDnew= PD/100; EDnew= ED/100; SHDInew= SHDI × 10; SHEInew= SHEI × 10; SIDInew=
SIDI × 10; SIEInew= SIEI × 10. Each model has three independent variables, including buffer radius, the percentage of forest area and a landscape metric.
67
revealed positive associations between BU with several land-scape fragmentation and diversity metrics.
In this study, we used metrics to quantify landscape diver-sity for each village and found that they were significantly different between BU case villages and reference villages at the 4 km extent. The results suggested that landscape con-text could be used to predict BU presence at the village level. It follows that these metrics can be used to prioritize the location of conservation and health treatment activities, independent of the availability of disease data, which is often limited and uncertain in rural areas.
It should be cautioned that the BU cases used in this study were collected from hospitals and clinics only. There-fore, patients who did not visit hospitals or clinics were not recorded in our dataset. In addition, BU cases were con-firmed through symptoms but not laboratory tests. It cannot completely rule out false-positive patients. As a result, our BU cases might be either underestimated or overestimated.
In summary, our analyses demonstrated the connection between land cover disturbance and BU disease in south-western Ghana. Specifically, we found that there were sig-nificant differences in landscape pattern metrics between BU case villages and reference villages, suggesting that the increased fragmentation and diversity of landscape structure may be potential risk factors for the emergence of BU dis-ease. Understanding these connections may provide insights to reduce the risk of BU disease in Ghana, and other areas experiencing rapid land cover change and provide a metric for prioritizing conservation and health activities in data-poor regions.
Received September 4, 2015. Accepted for publication March 17, 2016.
Published online May 16, 2016.
Note: Supplemental figures and tables appear at www.ajtmh.org.
Acknowledgments: We would like to acknowledge our other team members in the United States and in Ghana for their assistance.
Financial support: The study was supported by an NSF CNH grant (no. 0909447).
Authors’ addresses: Jianyong Wu, Gillings School of Global Public Health, The University of North Carolina, Chapel Hill, NC, E-mail: [email protected]. Erica A. H. Smithwick, Department of Geography, The Pennsylvania State University, University Park, PA, E-mail: [email protected].
REFERENCES
1. Turner MG, 2010. Disturbance and landscape dynamics in a changing world.Ecology 91:2833–2849.
2. Molyneux D, Ostfeld RS, Bernstein A, Chivian E, 2008. Ecosys-tem disturbance, biodiversity loss, and human infectious dis-ease. In:In Sustaining Life: How Human Health Depends on Biodiversity.E Chivian, A Bernstein, eds. Oxford, New York: Oxford Univ. Press, 287–323.
3. Patz JA, Daszak P, Tabor GM, Aguirre AA, Pearl M, Epstein J, Wolfe ND, Kilpatrick AM, Foufopoulos J, Molyneux D, Bradley DJ; Working Group on Land Use Change and Disease Emergence, 2004. Unhealthy landscapes: policy rec-ommendations on land use change and infectious disease emergence.Environ Health Perspect 112:1092–1098.
4. Vittor AY, Pan W, Gilman RH, Tielsch J, Glass G, Shields T, Sanchez-Lozano W, Pinedo VV, Salas-Cobos E, Flores S, Patz JA, 2009. Linking deforestation to malaria in the Amazon: characterization of the breeding habitat of the principal malaria vector,Anopheles darlingi.Am J Trop Med Hyg 81:5–12.
5. Southgate VR, 1997. Schistosomiasis in the Senegal River Basin: before and after the construction of the dams at Diama, Senegal and Manantali, Mali and future prospects.J Helminthol 71:125–132.
6. Schmidt KA, Ostfeld RS, 2001. Biodiversity and the dilution effect in disease ecology.Ecology 82:609–619.
7. Campbell LP, Finley AO, Benbow ME, Gronseth J, Small P, Johnson RC, Sopoh GE, Merritt RM, Williamson H, Qi J, 2015. Spatial analysis of anthropogenic landscape distur-bance and Buruli ulcer disease in Benin. PLoS Negl Trop Dis 9:e0004123.
8. Brou T, Broutin H, Elguero E, Asse H, Guegan JF, 2008. Land-scape diversity related to Buruli ulcer disease in Cote d’Ivoire. PLoS Negl Trop Dis 2:e271.
9. Wagner T, Benbow ME, Brenden TO, Qi J, Johnson RC, 2008. Buruli ulcer disease prevalence in Benin, west Africa: associa-tions with land use/cover and the identification of disease clusters.Int J Health Geogr 7:25.
10. Wu J, Tschakert P, Klutse E, Ferring D, Ricciardi V, Hausermann H, Oppong J, Smithwick EA, 2015. Buruli ulcer disease and its association with land cover in southwestern Ghana.PLoS Negl Trop Dis 9:e0003840.
11. Merritt RW, Benbow ME, Small PL, 2005. Unraveling an emerging disease associated with disturbed aquatic envi-ronments: the case of Buruli ulcer. Front Ecol Environ 3: 323–331.
12. Marsollier L, Robert R, Aubry J, Saint Andre JP, Kouakou H, Legras P, Manceau AL, Mahaza C, Carbonnelle B, 2002. Aquatic insects as a vector forMycobacterium ulcerans.Appl Environ Microbiol 68:4623–4628.
13. Benbow ME, Williamson H, Kimbirauskas R, McIntosh MD, Kolar R, Quaye C, Akpabey F, Boakye D, Small P, Merritt RW, 2008. Aquatic invertebrates as unlikely vectors of Buruli ulcer disease.Emerg Infect Dis 14:1247–1254.
14. Fahrig L, 2003. Effects of habitat fragmentation on biodiversity. Annu Rev Ecol Evol Syst 34:487–515.
15. WHO, 2013.Global Health Observatory (GHO)—Buruli Ulcer. Available at: http://www.who.int/gho/neglected_diseases/buruli_ ulcer/en/. Accessed March 12, 2016.
16. Hansen CP, Lund JF, Treue T, 2009. Neither fast, nor easy: the prospect of reduced emissions from deforestation and degra-dation (REDD) in Ghana.Int For Rev 11:439–455.
17. Breisinger C, Diao X, Thurlow J, Al-Hassan RM, 2008.Agriculture for Development in Ghana: New Opportunities and Challenges. Washington, DC: International Food Policy Research Institute. 18. Bloch R, Owusu G, 2012. Linkages in Ghana’s gold mining
industry: challenging the enclave thesis.Resources Policy 37: 434–442.
19. Hilson G, Clifford MJ, 2010. Small-scale gold mining, the envi-ronment and human health: an introduction to the Ghana case.Int J Environ Pollut 41:185–194.
20. York AM, Shrestha M, Boone CG, Zhang SA, Harrington JA, Prebyl TJ, Swann A, Agar M, Antolin MF, Nolen B, Wright JB, Skaggs R, 2011. Land fragmentation under rapid urbani-zation: a cross-site analysis of southwestern cities. Urban Ecosyst 14:429–455.
21. McGarigal K, Cushman S, Ene E, 2012. FRAGSTATS v4: Spa-tial Pattern Analysis Program for Categorical and Continuous Maps. Computer Software Program Produced by the Authors at the University of Massachusetts, Amherst, MA.
22. Gelman A, 2008. Scaling regression inputs by dividing by two standard deviations.Stat Med 27:2865–2873.
23. Ostfeld RS, Keesing F, 2000. Biodiversity and disease risk: the case of Lyme disease.Conserv Biol 14:722–728.
24. Allan BF, Keesing F, Ostfeld RS, 2003. Effect of forest fragmen-tation on Lyme disease risk.Conserv Biol 17:267–272. 25. Brownstein JS, Skelly DK, Holford TR, Fish D, 2005. Forest
fragmentation predicts local scale heterogeneity of Lyme disease risk.Oecologia 146:469–475.
26. Langlois JP, Fahrig L, Merriam G, Artsob H, 2001. Landscape structure influences continental distribution of hantavirus in deer mice.Landsc Ecol 16:255–266.
on hantavirus prevalence in Panama.Ann N Y Acad Sci 1149: 80–83.
28. Levin SA, 1992. The problem of pattern and scale in ecol-ogy: the Robert H. MacArthur award lecture.Ecology 73: 1943–1967.
29. Collinge SK, Johnson WC, Ray C, Matchett R, Grensten J, Cully JF Jr, Gage KL, Kosoy MY, Loye JE, Martin AP, 2005.
Landscape structure and plague occurrence in black-tailed prairie dogs on grasslands of the western USA.Landsc Ecol 20:941–955.
30. Wagner T, Benbow ME, Burns M, Johnson RC, Merritt RW, Qi J, Small PL, 2008. A landscape-based model for predicting Mycobacterium ulceransinfection (Buruli ulcer disease) pres-ence in Benin, west Africa.EcoHealth 5:69–79.
69
